UAV image small target detection network via attention mechanism and multi-scale feature fusion

Aiming at the challenges of small-target feature extraction and low detection accuracy in object detection caused by dense small targets and large target scale variations in UAV aerial images, this paper proposes an Aerial Small-Target Detection Network with Attention Mechanism and Multi-Scale Feature Fusion (AMSFN). First, a Multi-Scale Dilated Fusion Module (MDFM) is designed to enhance the network's feature extraction capability through the combination of multi-scale feature fusion and attention mechanisms. Second, a Global Grouped Coordinate Attention Module (GGCA) is developed to capture multi-dimensional global information and strengthen the feature representation of small targets. Then, the Normalized Wasserstein Distance (NWD) loss function is combined with the Complete Intersection over Union (CIOU) loss function as the localization regression loss to accelerate network convergence. Finally, an additional target detection layer is introduced, and K-means++ is used to cluster anchor boxes. Experimental results on the VisDrone2019 dataset show that compared with YOLOv7, AMSFN improves mAP0.5 and mAP0.5:0.95 by 5.2% and 4.3%, respectively, significantly enhancing the detection accuracy of small targets.